Data driven hyperparameter optimization of one-class support vector machines for anomaly detection in wireless sensor networks

Abstract

One-class support vector machines (OCSVM) have been recently applied to detect anomalies in wireless sensor networks (WSNs). Typically, OCSVM is kernelized by radial bais functions (RBF, or Gausian kernel) whereas selecting Gaussian kernel hyperparameter is based upon availability of anomalies, which is rarely applicable in practice. This article investigates the application of OCSVM to detect anomalies in WSNs with data-driven hyperparameter optimization. Specifically, the information of the farthest and the nearest neighbors of each sample is used to construct the objective cost instead of labeling based metrics such as geometric mean accuracy (G-mean) or area under the receiver operating characteristic (AUROC). The efficiency of this method is illustrated over the IBRL dataset whereas the resulting estimated boundary as well as anomaly detection performance are comparable with existing methods.